A machine learning pipeline is an automated, repeatable workflow that takes data from source to a deployed, monitored model: ingest and validate data, engineer features, train, evaluate, register, deploy and monitor. In 2026 the same pipeline idea also covers LLM and agent workloads, adding steps for prompt and retrieval versioning, LLM evaluation and tracing.
Updated October 2026
What changed in 2026
- Pipelines now handle GenAI. MLflow 3, launched in June 2025, added tracing and GenAI evaluation alongside classic model tracking, so one pipeline can version prompts, retrieval logic and models together.
- Agents are moving into production. McKinsey's 2026 State of AI reports that about two in ten respondents' organisations have scaled AI agents enterprise-wide, and agents need the same release discipline as models.
- Cloud pipeline services are mature and standardised: Azure Machine Learning pipelines, SageMaker Pipelines and Vertex AI Pipelines all follow the same step-based model.
- Regulation adds documentation steps. The EU AI Act's high-risk obligations now apply from 2 December 2027 for stand-alone systems, and pipelines are the natural place to record data lineage and test results.
Why ML pipelines matter
Training a model in a notebook is the easy part. Google's well-known paper on hidden technical debt in ML systems showed that model code is a small fraction of a real ML system; data collection, validation, serving and monitoring make up most of it. A pipeline turns all of that into code that runs the same way every time, so results are reproducible, reviews are possible and retraining is routine.
The stages of a machine learning pipeline
| Stage | What happens | Typical 2026 tools |
|---|---|---|
| 1. Ingest | Pull data from databases, files, streams and devices | Spark, Kafka, Airflow, cloud connectors |
| 2. Validate | Check schema, ranges, missing values and drift | Great Expectations, TFX Data Validation |
| 3. Transform and features | Clean, join, encode, store reusable features | Pandas, Spark, Feast or cloud feature stores |
| 4. Train | Fit models, tune hyperparameters, track runs | PyTorch, scikit-learn, XGBoost, MLflow |
| 5. Evaluate | Test on held-out data, slices, fairness and robustness | MLflow evaluate, custom test suites |
| 6. Register | Version the model with lineage and approval status | Model registry (MLflow, SageMaker, Vertex, Azure ML) |
| 7. Deploy | Batch, real-time API, streaming or edge | KServe, vLLM, Triton, edge runtimes |
| 8. Monitor and retrain | Track accuracy, drift, latency, cost; trigger retraining | Evidently, Prometheus, cloud monitors |
Google's ML pipelines guide splits this into data, training, validation and serving pipelines, which is a helpful way to assign ownership.
Pipeline architecture patterns
Batch training, batch scoring
Retrain nightly or weekly and score in bulk, for example demand forecasts. Simple and cheap.
Batch training, online serving
Retrain on a schedule; serve predictions through an API in milliseconds. The most common pattern for fraud, recommendations and pricing.
Continuous training
Monitoring triggers retraining when drift crosses a threshold. Needs strong automated tests so a bad model never ships.
Edge pipelines
Train centrally, compress the model and push it to thousands of devices, then collect feedback. See MLOps for edge AI and TinyML and MLOps.
LLM and agent pipelines
Instead of training from scratch, the pipeline versions the base model, prompts, retrieval index and tools, then runs an evaluation suite (accuracy, groundedness, safety) before release. Fine-tuning and distributed training slot in when needed; see distributed machine learning.
Best practices
- Start with the business metric and the decision the model drives.
- Validate data before training, not after a bad model ships.
- Version everything: data snapshots, code, features, models, prompts and configs.
- Automate tests as gates in CI/CD, including slice and fairness checks. See ML model testing.
- Separate environments (dev, staging, prod) with promotion through the registry.
- Monitor in production for drift, latency and cost, and define who acts on alerts.
- Keep pipelines portable using containers and open formats, so you are not locked to one cloud.
Common mistakes
- Training-serving skew: features computed differently in training and production.
- Leakage: future information sneaking into training data.
- No owner for monitoring alerts, so drift goes unnoticed.
- Treating LLM apps as "no pipeline needed" because nothing is trained.
Choosing an orchestrator
| Option | Good fit | Watch out for |
|---|---|---|
| Kubeflow Pipelines | Kubernetes-native teams that want open source and portability | Cluster operations effort |
| Apache Airflow | Data teams already scheduling ETL jobs | Not ML-aware by default; add tracking separately |
| SageMaker, Vertex AI or Azure ML pipelines | Teams standardised on one cloud | Lock-in; harder to run on-premises or at the edge |
| Databricks workflows with MLflow | Lakehouse-centred data and ML teams | Edge and air-gapped deployment needs extra work |
The right answer often mixes these: a cloud service for training, an open-source registry for portability and a separate runtime for on-premises or edge serving. IBM's overview of the machine learning pipeline is a useful neutral reference when comparing vendors. Whatever you choose, keep pipeline definitions in version control and make every step runnable on its own, so failures can be retried without rerunning the whole flow.
Where it's heading: 2026 → 2030
| Period | ML pipelines |
|---|---|
| 2020 (history) | Notebook-to-production handoffs; early Kubeflow and TFX adoption |
| 2026 (now) | One pipeline for classic ML, LLMs and agents; tracing and GenAI evaluation built in |
| 2028 (expected) | Compliance evidence (lineage, test reports) generated automatically for regulated AI |
| 2030 (our view) | Agents help run pipelines: proposing retraining, writing tests and flagging drift for human approval |
How NexaStack helps
NexaStack provides the infrastructure layer for ML pipelines on your own cloud, data centre or edge: GPU scheduling, model serving, edge rollout and monitoring for classic models, LLMs and agents. Teams keep their preferred pipeline tools and use NexaStack to run them consistently wherever the data lives. For a comparison of the tools around it, see ML platform tools.
FAQ
What is a machine learning pipeline?
An automated workflow that moves data through preparation, training, evaluation, deployment and monitoring so models can be built and updated reliably.
What are the stages of a machine learning pipeline?
Ingest, validate, transform and engineer features, train, evaluate, register, deploy, and monitor and retrain.
How do you build a machine learning pipeline?
Define the goal and metric, script each stage as a reusable step, orchestrate them with a tool such as Kubeflow, Airflow or a cloud pipeline service, add tests as gates and connect monitoring to retraining.
What is a pipeline in LLMs?
An LLM pipeline versions prompts, retrieval data, model choice and tools, runs evaluations, deploys the app and traces its behaviour in production.
What are the latest ML pipeline trends in 2026?
Unified pipelines for ML and GenAI, built-in tracing and evaluation, edge rollout, compliance evidence and portable, multi-cloud deployment.
Run your ML pipelines anywhere
Need pipelines that train in one place and serve on-premises or at the edge? Book a NexaStack demo to see model serving, rollout and monitoring on your own infrastructure.
